ci: pack the native suites into four groups

[MarcusKainth]

Stacked on #591 (ci/driver-native-group); the diff against that branch is this one commit.

What this changes, and why

The six simulation groups were sized when a single test took 14 to 21 minutes. With every arm of a suite run in one session and the analysis cut, the slowest test now takes 227 s at its worst. The six groups finish their suites in 78 to 261 s each over the main runs 36126756550 to 36132815983, and emulator takes 210 to 246 s. Each group is a runner, and the runners queue against the account's concurrent-job limit.

The native crate's suites are repacked into four groups:

• The packing uses each test's worst time across those six runs.
• Each group is modelled as longest test first over four slots.
• Suites are placed greedily, then moved and swapped until the worst group stops falling.
• The crate's non-simulation suites stay in the catch-all group, native-sim-d, because group_coverage.rs only accepts simulation suites in the lettered lists.

The timeouts are re-derived from the same runs:

• nextest cutoff: 2,400 s drops to 720 s (period = "120s", terminate-after = 6). The slowest test's worst is 227 s. The cutoff sits above FIRST_TIC_TIMEOUT (600 s) plus a load, so a first tic that never comes fails with the session's own message.
• timeout-minutes for test and test-sim: 45 drops to 15. That is about twice the slowest test job seen (7.1 min, native-rest in 36131955884), and it stays above nextest's 12-minute cutoff.

The test jobs go from eight to six: emulator, driver-native and four simulation groups. ci-passed needs no change because the job names stay the same.

Evidence

model at 4 threads    tests   median   worst   sum of worst
native-sim-a            28     223 s   253 s     1,011 s
native-sim-b            26     219 s   251 s     1,000 s
native-sim-c            25     224 s   252 s     1,000 s
native-sim-d (+native) 399     232 s   253 s     1,013 s

alternatives, worst-time model: 3 groups 337 s, 4 groups 253 s, 5 groups 227 s (bound by sim_plat_live's longest test)

The same model against each group's observed suites step in the six runs (48 group-runs) gives observed/model between 0.96 and 1.17, mostly 1.05 to 1.10.

$ CARGO_PROFILE_TEST_DEBUG=0 scripts/test-group.sh --check; echo "check exit=$?"
emulator:      467 tests
native-sim-a:       28 tests
native-sim-b:       26 tests
native-sim-c:       25 tests
native-sim-d:      399 tests
driver-native:       15 tests
every test:      960; selected:      960
check exit=0
$ cargo test --locked -p clickdoom-native --test group_coverage
test result: ok. 2 passed; 0 failed
$ shellcheck scripts/test-group.sh; make actionlint zizmor; cargo nextest list -p clickdoom-native --test group_coverage
all exit 0 (the last parses the edited .config/nextest.toml)

What this PR's CI run should show

• Simulation groups: native-sim-a to native-sim-d each run their suites step in about 235 to 280 s, which is the model times 1.05 to 1.10 plus spread. Before, the six groups took 78 to 261 s.
• driver-native: about 215 s, as ci: run the driver's native suites as their own group #591's first run measured.
• emulator: 210 to 250 s, unchanged.
• Critical path: the slowest simulation group, about 20 to 30 s longer than emulator, with two fewer runners.
• test-groups: 960 of 960, with the counts above.

If the simulation groups come in past about 280 s, five groups (modelled at 227 s) is the next step.

Invariants

None. Test grouping and timeouts only.

Spec impact

• None. No contract in SPEC.md is touched

Checks

• make gates: not run. The checks above ran locally
• No AI attribution trailers in the commits

Written mostly by Claude Opus 5.5.